Patient Re-Identification Based on Deep Metric Learning in Trunk Computed Tomography Images Acquired from Devices...

Area of Science:

• Medical Imaging
• Radiology
• Computer Vision

Background:

• Radiologists rely on image metadata for patient identification during interpretation.
• Incorrect patient identifiers in medical images pose a significant challenge, potentially leading to diagnostic errors.

Purpose of the Study:

• To develop and evaluate a patient re-identification technique for verifying and correcting metadata associated with computed tomography (CT) images.
• To improve the accuracy of patient identification in radiological workflows.

Main Methods:

• A feature vector matching technique using a deep feature extractor was employed to link correct metadata to CT image sets.
• The method calculates similarity scores between baseline and follow-up images to identify correct patient metadata.
• An augmentation technique was developed to enhance re-identification performance across different vendors and imaging variations.

Main Results:

• The proposed method successfully linked correct metadata to CT images with lost or wrongly assigned identifiers.
• Similarity scores were significantly higher for correctly matched patient images compared to incorrectly matched ones.
• The deep feature extractor demonstrated robustness in distinguishing patients across different vendors without additional training.

Conclusions:

• The developed patient re-identification method effectively detects incorrect patient identifiers in medical imaging metadata.
• This approach can help mitigate diagnostic errors stemming from human error or technical issues in data assignment.
• The technique offers a reliable solution for ensuring data integrity in radiological interpretations.